>the first couple of KB. Given the loader does so little, it hardly seems
>worth bothering to run it. You need to nkow the h/w to drive it, and a
>paging or reset register isn't exactly difficult.
Yeah, that's my feeling too.
>Why bother with loaders, at the expense of thread switch times? On
>NetBSD, I shortcut switching between different kernel threads (eg
>process reaper) although that patch hasn't been committed yet which I
>think is kind-of what Phil suggested.
Basically. Since a kernel thread (usually) doesn't have any memory mapping in
the area normally touched by user tasks, we don't have to worry about cache
issues. In fact I've now thought of a way to play things that will hopefully
allow us to handle the case where a user process is fighting a kernel thread
for the CPU without any cache flush at all, and the clone() case comes out in
the wash which is an extra bonus. :-)
Essentially what you do is just to remember the last set of pagetables you
loaded for a non-kernel thread (having nobbled kecardd such that it doesn't
count as a kernel thread any more). If it's the same as the one you're about
to load in, or if the new task is a kernel thread, you can skip the cache and
TLB drama.
I'll see if I can come up with a patch later. By the way, before I send the
CATS board back, do you guys have any benchmark numbers for NetBSD running on
one of those? I'd be interested to see how Linux compares.
p.
unsubscribe: body of `unsubscribe linux-arm' to [EMAIL PROTECTED]
